Penn1357/bilibili-clickbait-xlmr

bilibili-clickbait-xlmr

XLM-RoBERTa-base fine-tuned as a single-logit sigmoid regressor to score Chinese video titles for clickbait ("标题党"). Built for bilibili-clickbait-filter, a fork of festoney8/bilibili-cleaner that replaces keyword-based filtering with a distilled model.

Training

Distilled from an averaged two-teacher signal — codex (gpt-5.6-sol, AUC 0.909 on the eval set below) and Gemini 3.1 Pro (AUC 0.913) — soft-labeling 10,684 bilibili video titles scraped from public ranking/popular/newlist/search APIs. Averaging the two teachers measured AUC 0.930, meaningfully above either alone (correlation 0.81 — independent enough that averaging helps). Fine-tuned on the averaged labels; checkpoint selected by validation loss, not test AUC.

Evaluation

200 titles, hand-labeled by a single human annotator, held out from every training/distillation step:

Full writeup, ablations (data scale vs. teacher quality — the latter is what actually helped), and honest failure modes: see the project README.

Usage

import { AutoTokenizer, AutoModelForSequenceClassification } from '@huggingface/transformers'

const tok = await AutoTokenizer.from_pretrained('Penn1357/bilibili-clickbait-xlmr')
const model = await AutoModelForSequenceClassification.from_pretrained('Penn1357/bilibili-clickbait-xlmr', { dtype: 'fp32' })

const inputs = await tok(['震惊！99%的人都不知道的秘密'], { padding: true, truncation: true })
const { logits } = await model(inputs)
const score = 1 / (1 + Math.exp(-logits.data[0]))  // sigmoid -- this is a regressor, not a softmax classifier

Not quantized (fp32, ~1.1GB) by request — this is the tradeoff of full precision vs. download size and in-browser inference speed.